Role and Persona Prompting: Making AI Sound Like Someone Specific, Part 1

AI gives generic answers because it does not know what role to play. Tell it. 'Act as a fitness coach.' 'Act as a kind older sibling.' 'Act as my study buddy.' The answers change based on the role.

Some examples

• For homework help: 'Act as a patient tutor for a 9th grader. Explain step by step.'
• For tough feelings: 'Act as a kind, calm friend who listens before giving advice.'
• For project planning: 'Act as a project manager helping me break this into steps.'
• For practice: 'Act as my interviewer for a job at a coffee shop. Ask me questions.'

Try it!

Long descriptions can confuse AI. Real examples are way clearer. 'Make it like this [example]' beats 'make it formal but not too formal but also friendly.'

Some examples

• Need an essay opener? Give AI an opener you like and say 'do something like this.'
• Need a poem in a specific style? Give it a poem you like.
• Need a code style? Show AI an example of the style you want.
• Need a tone for an email? Show a sample email.

Try it!

Understanding "Real Examples Beat Long Descriptions Every Time" in practice: Prompting is a skill: the more specific and structured your input, the more useful the output. If you want a specific style, give AI an example of what you mean. Way better than describing it — and knowing how to apply this gives you a concrete advantage.

• Apply examples in your prompting workflow to get better results
• Apply few-shot prompting in your prompting workflow to get better results
• Apply style matching in your prompting workflow to get better results

1. 1Rewrite one of your best prompts using role + context + task + format
2. 2Ask an AI to critique your prompt and suggest improvements
3. 3Compare outputs from two models using the same prompt

'You are a tutor' is a role. 'You are a patient tutor who never makes me feel dumb and uses skateboarding metaphors' is a persona. The second one gives you way better answers because AI now has a voice to hold.

Some examples

• 'Be a coach who's encouraging but never sugar-coats.'
• 'Talk like a sarcastic older sibling who actually wants me to learn.'
• 'You're an editor who hates flowery writing.'
• 'Be a Gen Z science teacher who explains things using TikTok comparisons.'

Try it!

Ask AI to give you feedback on a paragraph as 'a writing teacher.' Then ask again as 'a tired but funny editor who's seen 1,000 essays this week.' Compare the energy.

Tone locking means giving AI clear rules about how to sound: word length, formality, slang, punchiness. Without it, AI defaults to bland 'I'd be happy to help!' tone that no one wants to read.

Some examples

• Bad: 'Write a caption.' Better: 'Write a caption. Casual, no exclamation marks, max 12 words.'
• Tell AI: 'Sound like you're texting a friend, not pitching a startup.'
• Give an example of your real writing and say 'Match this voice.'
• Ban specific words: 'Never use *delve*, *unleash*, or *journey*.'

Try it!

Write the same prompt twice — once vague, once with a strict tone lock. Compare and notice the difference.

Roleplay prompts assign AI a persona so it answers from that point of view. A 'pediatrician' will write a different first-aid answer than a 'random helpful assistant.'

Some examples

• 'Act as my high school English teacher grading this essay strictly.'
• 'Be a startup founder who's seen 100 pitches. Tear mine apart.'
• 'Roleplay a parent with anxiety about their kid going to college.'
• 'You're a friend who hates BS. Tell me if this idea is dumb.'

Try it!

Take a question you have and pick a persona for AI to answer as. Compare it to a generic answer. Notice the angle.

Stuck on tone? Ask AI for 5 versions: formal, casual, hype, sarcastic, soft. Pick your favorite or remix two. Tone ladders cure 'I don't know how to say it.'

Some examples

• Prompt: 'Write this in 5 tones: corporate, Gen Z, professor, your-grandma, sportscaster.'
• Tone ladders help when texting a teacher or boss for the first time.
• Pair tone ladders with anchor examples for laser-targeted voice.
• Save the best tone variant as a template for next time.

Try it!

Write one sentence — an apology, a pitch, anything. Have AI rewrite it in 5 tones. Pick a winner.

'You are a world-class chef' makes ChatGPT use chef vocabulary. It does NOT teach it new recipes. Role prompts are great for tone and vocabulary, useless for facts and skill. The 2026 evidence is clear: 'expert' tags don't unlock hidden knowledge. Use them for voice, not for IQ.

Some examples

• Useful: 'You are a fitness coach. Reply in 3 short bullets.' (sets tone & format)
• Useful: 'You are a strict English teacher. Mark grammar errors.' (sets job & rigor)
• Useless: 'You are a Nobel-prize-winning physicist.' (won't make the answer more correct)
• Useless: 'You are the world's best programmer.' (still hallucinates the same APIs)

Try it!

Pick one role-prompt you use a lot. Re-ask the same question without the role and compare. Notice what actually changed.

Role-prompting is the trick of giving the AI a job before the task. 'You are a code reviewer' makes it pickier than 'tell me what's wrong with this code.' The role pulls the model's training toward responses that fit that role.

Some examples

• 'You are a TikTok scriptwriter' — punchier, hooks first.
• 'You are a college admissions reader' — sharper feedback on your essay.
• 'You are a senior backend engineer' — Claude flags performance issues a generic prompt misses.
• 'You are a 12-year-old' — explanations get simpler, analogies get sillier.

Try it!

Take a prompt you used recently. Add 'You are a [specific expert]' at the top. Run both versions. Compare.

Anthropic specifically trained Claude to recognize XML tags as section dividers. Wrapping your prompt's parts — task, context, examples, output format — in tags makes it more reliable. It's like headers in a document for the AI to follow.

Some examples

• `<task>Summarize this article</task><article>...</article>` — clean separation.
• `<example>input → output</example>` blocks for few-shot prompts.
• `<rules>Don't use emojis. Cite sources.</rules>` keeps your constraints from blending into the task.
• `<output_format>JSON only, no prose</output_format>` makes the format ask stand out.

Try it!

Take a multi-part prompt you wrote. Wrap each part in XML tags. Run both versions in Claude. Compare.